Part-aware primitive learning for efficient 3D shape completion

Point cloud completion aims to infer complete 3D geometry from partial scans. Most existing approaches rely on dense point-based decoding or large transformer models that implicitly learn geometry and overlook the compositional nature of man-made objects. We propose PPCNet, a part-aware primitive completion network that reconstructs shapes through explicit geometric abstraction and relational refinement. The model decomposes a partial scan into a compact set of learnable superquadric primitives, refines their spatial organization via geometry-aware graph attention, and adaptively recovers point-level details with a lightweight residual generator. Experiments on the PCN dataset show that PPCNet achieves competitive reconstruction quality with 4 × fewer FLOPs than transformer baselines. Beyond empirical performance, its primitive-based design enhances interpretability and structural stability, highlighting the potential of explicit geometric priors for efficient and controllable 3D completion.